Model experiments
Pruning LFM2: smaller weights, measured limits
Reducing an MoE model’s expert count produced a smaller runtime artifact and a modest latency difference in two small historical runs.
- Experts retained
- 64 → 56
- Storage saved
- 1.61 GiB
- Mean job time · small source-PC runs
- 4.495 → 4.006 s
The question
Can reducing an MoE model's expert count lower its storage cost while retaining basic response behavior?
How the work was approached
Static router norm/bias scoring retained 56 of 64 experts. Compared original/pruned models using saved seven-prompt smoke checks and ten-job conveyor runs.
What the record shows
The prune summary records 1.61 GiB saved. Recomputed raw CSV means match the report: average end-to-end job time fell from 4.495 to 4.006 seconds in these two small runs.
What remains unresolved
Single historical runs with slightly different completion-token totals (582 vs 587); no independent held-out semantic quality assessment or causal performance guarantee. Read raw instruction and coding responses as a spot check; no generated code executed. No R640 evidence.
What we are trying to achieve
We want to reduce the memory and storage cost of local AI without quietly losing the behaviors that make it useful. The smaller artifact and these short-run timings are an early step. A stronger study would compare independent task quality, repeatability, memory use, and total latency under matched conditions.
Sources & reading notes
- LFM2_24B_A2B_TUNING.mdPhase 3 Pruning Retained project record Source record retained privately.
- prune-summary-56x.mdOutput sizes and dropped expert indices Retained project record Source record retained privately.
- prune-plan-56x.jsonExact retained/dropped index mapping Retained project record Source record retained privately.
- results.csvOriginal raw jobs, recomputed Retained project record Source record retained privately.
- results.csvPruned raw jobs, recomputed Retained project record Source record retained privately.
- prune-lfm2-gguf-56x.pyStatic norm/bias keep-plan computation Retained project record Source record retained privately.
This is a dated, source-based research account, not a live machine report. Historical results, engineering preparation, research designs, open review gates, and measured pilots remain distinct.